[-next,v2] mm/page_alloc: fix a false memory corruption
diff mbox series

Message ID 1561063566-16335-1-git-send-email-cai@lca.pw
State New
Headers show
Series
  • [-next,v2] mm/page_alloc: fix a false memory corruption
Related show

Commit Message

Qian Cai June 20, 2019, 8:46 p.m. UTC
The linux-next commit "mm: security: introduce init_on_alloc=1 and
init_on_free=1 boot options" [1] introduced a false positive when
init_on_free=1 and page_poison=on, due to the page_poison expects the
pattern 0xaa when allocating pages which were overwritten by
init_on_free=1 with 0.

Fix it by switching the order between kernel_init_free_pages() and
kernel_poison_pages() in free_pages_prepare().

[1] https://patchwork.kernel.org/patch/10999465/

Signed-off-by: Qian Cai <cai@lca.pw>
---

v2: After further debugging, the issue after switching order is likely a
    separate issue as clear_page() should not cause issues with future
    accesses.

 mm/page_alloc.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Comments

Kees Cook June 21, 2019, 1:01 a.m. UTC | #1
On Thu, Jun 20, 2019 at 04:46:06PM -0400, Qian Cai wrote:
> The linux-next commit "mm: security: introduce init_on_alloc=1 and
> init_on_free=1 boot options" [1] introduced a false positive when
> init_on_free=1 and page_poison=on, due to the page_poison expects the
> pattern 0xaa when allocating pages which were overwritten by
> init_on_free=1 with 0.
> 
> Fix it by switching the order between kernel_init_free_pages() and
> kernel_poison_pages() in free_pages_prepare().

Cool; this seems like the right approach. Alexander, what do you think?

Reviewed-by: Kees Cook <keescook@chromium.org>

-Kees

> 
> [1] https://patchwork.kernel.org/patch/10999465/
> 
> Signed-off-by: Qian Cai <cai@lca.pw>
> ---
> 
> v2: After further debugging, the issue after switching order is likely a
>     separate issue as clear_page() should not cause issues with future
>     accesses.
> 
>  mm/page_alloc.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 54dacf35d200..32bbd30c5f85 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1172,9 +1172,10 @@ static __always_inline bool free_pages_prepare(struct page *page,
>  					   PAGE_SIZE << order);
>  	}
>  	arch_free_page(page, order);
> -	kernel_poison_pages(page, 1 << order, 0);
>  	if (want_init_on_free())
>  		kernel_init_free_pages(page, 1 << order);
> +
> +	kernel_poison_pages(page, 1 << order, 0);
>  	if (debug_pagealloc_enabled())
>  		kernel_map_pages(page, 1 << order, 0);
>  
> -- 
> 1.8.3.1
>
Alexander Potapenko June 21, 2019, 10:39 a.m. UTC | #2
On Fri, Jun 21, 2019 at 3:01 AM Kees Cook <keescook@chromium.org> wrote:
>
> On Thu, Jun 20, 2019 at 04:46:06PM -0400, Qian Cai wrote:
> > The linux-next commit "mm: security: introduce init_on_alloc=1 and
> > init_on_free=1 boot options" [1] introduced a false positive when
> > init_on_free=1 and page_poison=on, due to the page_poison expects the
> > pattern 0xaa when allocating pages which were overwritten by
> > init_on_free=1 with 0.
> >
> > Fix it by switching the order between kernel_init_free_pages() and
> > kernel_poison_pages() in free_pages_prepare().
>
> Cool; this seems like the right approach. Alexander, what do you think?
Can using init_on_free together with page_poison bring any value at all?
Isn't it better to decide at boot time which of the two features we're
going to enable?

> Reviewed-by: Kees Cook <keescook@chromium.org>
>
> -Kees
>
> >
> > [1] https://patchwork.kernel.org/patch/10999465/
> >
> > Signed-off-by: Qian Cai <cai@lca.pw>
> > ---
> >
> > v2: After further debugging, the issue after switching order is likely a
> >     separate issue as clear_page() should not cause issues with future
> >     accesses.
> >
> >  mm/page_alloc.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 54dacf35d200..32bbd30c5f85 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -1172,9 +1172,10 @@ static __always_inline bool free_pages_prepare(struct page *page,
> >                                          PAGE_SIZE << order);
> >       }
> >       arch_free_page(page, order);
> > -     kernel_poison_pages(page, 1 << order, 0);
> >       if (want_init_on_free())
> >               kernel_init_free_pages(page, 1 << order);
> > +
> > +     kernel_poison_pages(page, 1 << order, 0);
> >       if (debug_pagealloc_enabled())
> >               kernel_map_pages(page, 1 << order, 0);
> >
> > --
> > 1.8.3.1
> >
>
> --
> Kees Cook
Qian Cai June 21, 2019, 12:26 p.m. UTC | #3
On Fri, 2019-06-21 at 12:39 +0200, Alexander Potapenko wrote:
> On Fri, Jun 21, 2019 at 3:01 AM Kees Cook <keescook@chromium.org> wrote:
> > 
> > On Thu, Jun 20, 2019 at 04:46:06PM -0400, Qian Cai wrote:
> > > The linux-next commit "mm: security: introduce init_on_alloc=1 and
> > > init_on_free=1 boot options" [1] introduced a false positive when
> > > init_on_free=1 and page_poison=on, due to the page_poison expects the
> > > pattern 0xaa when allocating pages which were overwritten by
> > > init_on_free=1 with 0.
> > > 
> > > Fix it by switching the order between kernel_init_free_pages() and
> > > kernel_poison_pages() in free_pages_prepare().
> > 
> > Cool; this seems like the right approach. Alexander, what do you think?
> 
> Can using init_on_free together with page_poison bring any value at all?
> Isn't it better to decide at boot time which of the two features we're
> going to enable?

I think the typical use case is people are using init_on_free=1, and then decide
to debug something by enabling page_poison=on. Definitely, don't want
init_on_free=1 to disable page_poison as the later has additional checking in
the allocation time to make sure that poison pattern set in the free time is
still there.

> 
> > Reviewed-by: Kees Cook <keescook@chromium.org>
> > 
> > -Kees
> > 
> > > 
> > > [1] https://patchwork.kernel.org/patch/10999465/
> > > 
> > > Signed-off-by: Qian Cai <cai@lca.pw>
> > > ---
> > > 
> > > v2: After further debugging, the issue after switching order is likely a
> > >     separate issue as clear_page() should not cause issues with future
> > >     accesses.
> > > 
> > >  mm/page_alloc.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 54dacf35d200..32bbd30c5f85 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -1172,9 +1172,10 @@ static __always_inline bool
> > > free_pages_prepare(struct page *page,
> > >                                          PAGE_SIZE << order);
> > >       }
> > >       arch_free_page(page, order);
> > > -     kernel_poison_pages(page, 1 << order, 0);
> > >       if (want_init_on_free())
> > >               kernel_init_free_pages(page, 1 << order);
> > > +
> > > +     kernel_poison_pages(page, 1 << order, 0);
> > >       if (debug_pagealloc_enabled())
> > >               kernel_map_pages(page, 1 << order, 0);
> > > 
> > > --
> > > 1.8.3.1
> > > 
> > 
> > --
> > Kees Cook
> 
> 
>
Alexander Potapenko June 21, 2019, 2:37 p.m. UTC | #4
On Fri, Jun 21, 2019 at 2:26 PM Qian Cai <cai@lca.pw> wrote:
>
> On Fri, 2019-06-21 at 12:39 +0200, Alexander Potapenko wrote:
> > On Fri, Jun 21, 2019 at 3:01 AM Kees Cook <keescook@chromium.org> wrote:
> > >
> > > On Thu, Jun 20, 2019 at 04:46:06PM -0400, Qian Cai wrote:
> > > > The linux-next commit "mm: security: introduce init_on_alloc=1 and
> > > > init_on_free=1 boot options" [1] introduced a false positive when
> > > > init_on_free=1 and page_poison=on, due to the page_poison expects the
> > > > pattern 0xaa when allocating pages which were overwritten by
> > > > init_on_free=1 with 0.
> > > >
> > > > Fix it by switching the order between kernel_init_free_pages() and
> > > > kernel_poison_pages() in free_pages_prepare().
> > >
> > > Cool; this seems like the right approach. Alexander, what do you think?
> >
> > Can using init_on_free together with page_poison bring any value at all?
> > Isn't it better to decide at boot time which of the two features we're
> > going to enable?
>
> I think the typical use case is people are using init_on_free=1, and then decide
> to debug something by enabling page_poison=on. Definitely, don't want
> init_on_free=1 to disable page_poison as the later has additional checking in
> the allocation time to make sure that poison pattern set in the free time is
> still there.
In addition to information lifetime reduction the idea of init_on_free
is to ensure the newly allocated objects have predictable contents.
Therefore it's handy (although not strictly necessary) to keep them
zero-initialized regardless of other boot-time flags.
Right now free_pages_prezeroed() relies on that, though this can be changed.

On the other hand, since page_poison already initializes freed memory,
we can probably make want_init_on_free() return false in that case to
avoid extra initialization.

Side note: if we make it possible to switch betwen 0x00 and 0xAA in
init_on_free mode, we can merge it with page_poison, performing the
initialization depending on a boot-time flag and doing heavyweight
checks under a separate config.

> >
> > > Reviewed-by: Kees Cook <keescook@chromium.org>
> > >
> > > -Kees
> > >
> > > >
> > > > [1] https://patchwork.kernel.org/patch/10999465/
> > > >
> > > > Signed-off-by: Qian Cai <cai@lca.pw>
> > > > ---
> > > >
> > > > v2: After further debugging, the issue after switching order is likely a
> > > >     separate issue as clear_page() should not cause issues with future
> > > >     accesses.
> > > >
> > > >  mm/page_alloc.c | 3 ++-
> > > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > > index 54dacf35d200..32bbd30c5f85 100644
> > > > --- a/mm/page_alloc.c
> > > > +++ b/mm/page_alloc.c
> > > > @@ -1172,9 +1172,10 @@ static __always_inline bool
> > > > free_pages_prepare(struct page *page,
> > > >                                          PAGE_SIZE << order);
> > > >       }
> > > >       arch_free_page(page, order);
> > > > -     kernel_poison_pages(page, 1 << order, 0);
> > > >       if (want_init_on_free())
> > > >               kernel_init_free_pages(page, 1 << order);
> > > > +
> > > > +     kernel_poison_pages(page, 1 << order, 0);
> > > >       if (debug_pagealloc_enabled())
> > > >               kernel_map_pages(page, 1 << order, 0);
> > > >
> > > > --
> > > > 1.8.3.1
> > > >
> > >
> > > --
> > > Kees Cook
> >
> >
> >
Qian Cai June 21, 2019, 2:56 p.m. UTC | #5
On Fri, 2019-06-21 at 16:37 +0200, Alexander Potapenko wrote:
> On Fri, Jun 21, 2019 at 2:26 PM Qian Cai <cai@lca.pw> wrote:
> > 
> > On Fri, 2019-06-21 at 12:39 +0200, Alexander Potapenko wrote:
> > > On Fri, Jun 21, 2019 at 3:01 AM Kees Cook <keescook@chromium.org> wrote:
> > > > 
> > > > On Thu, Jun 20, 2019 at 04:46:06PM -0400, Qian Cai wrote:
> > > > > The linux-next commit "mm: security: introduce init_on_alloc=1 and
> > > > > init_on_free=1 boot options" [1] introduced a false positive when
> > > > > init_on_free=1 and page_poison=on, due to the page_poison expects the
> > > > > pattern 0xaa when allocating pages which were overwritten by
> > > > > init_on_free=1 with 0.
> > > > > 
> > > > > Fix it by switching the order between kernel_init_free_pages() and
> > > > > kernel_poison_pages() in free_pages_prepare().
> > > > 
> > > > Cool; this seems like the right approach. Alexander, what do you think?
> > > 
> > > Can using init_on_free together with page_poison bring any value at all?
> > > Isn't it better to decide at boot time which of the two features we're
> > > going to enable?
> > 
> > I think the typical use case is people are using init_on_free=1, and then
> > decide
> > to debug something by enabling page_poison=on. Definitely, don't want
> > init_on_free=1 to disable page_poison as the later has additional checking
> > in
> > the allocation time to make sure that poison pattern set in the free time is
> > still there.
> 
> In addition to information lifetime reduction the idea of init_on_free
> is to ensure the newly allocated objects have predictable contents.
> Therefore it's handy (although not strictly necessary) to keep them
> zero-initialized regardless of other boot-time flags.
> Right now free_pages_prezeroed() relies on that, though this can be changed.
> 
> On the other hand, since page_poison already initializes freed memory,
> we can probably make want_init_on_free() return false in that case to
> avoid extra initialization.
> 
> Side note: if we make it possible to switch betwen 0x00 and 0xAA in
> init_on_free mode, we can merge it with page_poison, performing the
> initialization depending on a boot-time flag and doing heavyweight
> checks under a separate config.

Yes, that would be great which will reduce code duplication.
Alexander Potapenko June 21, 2019, 3:26 p.m. UTC | #6
On Fri, Jun 21, 2019 at 4:56 PM Qian Cai <cai@lca.pw> wrote:
>
> On Fri, 2019-06-21 at 16:37 +0200, Alexander Potapenko wrote:
> > On Fri, Jun 21, 2019 at 2:26 PM Qian Cai <cai@lca.pw> wrote:
> > >
> > > On Fri, 2019-06-21 at 12:39 +0200, Alexander Potapenko wrote:
> > > > On Fri, Jun 21, 2019 at 3:01 AM Kees Cook <keescook@chromium.org> wrote:
> > > > >
> > > > > On Thu, Jun 20, 2019 at 04:46:06PM -0400, Qian Cai wrote:
> > > > > > The linux-next commit "mm: security: introduce init_on_alloc=1 and
> > > > > > init_on_free=1 boot options" [1] introduced a false positive when
> > > > > > init_on_free=1 and page_poison=on, due to the page_poison expects the
> > > > > > pattern 0xaa when allocating pages which were overwritten by
> > > > > > init_on_free=1 with 0.
> > > > > >
> > > > > > Fix it by switching the order between kernel_init_free_pages() and
> > > > > > kernel_poison_pages() in free_pages_prepare().
> > > > >
> > > > > Cool; this seems like the right approach. Alexander, what do you think?
> > > >
> > > > Can using init_on_free together with page_poison bring any value at all?
> > > > Isn't it better to decide at boot time which of the two features we're
> > > > going to enable?
> > >
> > > I think the typical use case is people are using init_on_free=1, and then
> > > decide
> > > to debug something by enabling page_poison=on. Definitely, don't want
> > > init_on_free=1 to disable page_poison as the later has additional checking
> > > in
> > > the allocation time to make sure that poison pattern set in the free time is
> > > still there.
> >
> > In addition to information lifetime reduction the idea of init_on_free
> > is to ensure the newly allocated objects have predictable contents.
> > Therefore it's handy (although not strictly necessary) to keep them
> > zero-initialized regardless of other boot-time flags.
> > Right now free_pages_prezeroed() relies on that, though this can be changed.
> >
> > On the other hand, since page_poison already initializes freed memory,
> > we can probably make want_init_on_free() return false in that case to
> > avoid extra initialization.
> >
> > Side note: if we make it possible to switch betwen 0x00 and 0xAA in
> > init_on_free mode, we can merge it with page_poison, performing the
> > initialization depending on a boot-time flag and doing heavyweight
> > checks under a separate config.
>
> Yes, that would be great which will reduce code duplication.
I suggest we disable init_on_alloc/init_on_free under
CONFIG_PAGE_POISONING now then and work towards deduplicating this
code in further patch series.

Patch
diff mbox series

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 54dacf35d200..32bbd30c5f85 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1172,9 +1172,10 @@  static __always_inline bool free_pages_prepare(struct page *page,
 					   PAGE_SIZE << order);
 	}
 	arch_free_page(page, order);
-	kernel_poison_pages(page, 1 << order, 0);
 	if (want_init_on_free())
 		kernel_init_free_pages(page, 1 << order);
+
+	kernel_poison_pages(page, 1 << order, 0);
 	if (debug_pagealloc_enabled())
 		kernel_map_pages(page, 1 << order, 0);